Experimental
Asset Similarity
Find visually related images and near-duplicate variants without sending your artwork to a cloud search service.
Two useful kinds of “similar”
- Similar: ranks compatible image embeddings to surface related subjects, compositions, and visual styles.
- Near duplicates: uses exact identity and visual fingerprints for resized, recompressed, or lightly edited copies.
Local and explicit
The current workflow runs image-to-image or text-to-image inference in a worker and scans up to 30 images in the open document. The optional image and text models are downloaded only after you ask for them, verified before use, and cached locally. It does not yet build a persistent cross-project library index.
Set the right expectation
Similarity scores are ranking signals, not truth. Results can favor color, background, style, or composition in surprising ways. Review candidates before using them in a workflow; Varve never deletes assets automatically because a model score is high.